Академический Документы
Профессиональный Документы
Культура Документы
Benjamin Durakovic
Industrial Engineering, International University of Sarajevo
1. Introduction
Design of Experiments (DOE) mathematical methodology used for planning and conducting experiments as
well as analyzing and interpreting data obtained from the experiments. It is a branch of applied statistics that is
used for conducting scientific studies of a system, process or product in which input variables (Xs) were
manipulated to investigate its effects on measured response variable (Y).
Over past two decades, DOE was a very useful tool traditionally used for improvement of product quality and
reliability [1]. The usage of DOE has been expanded across many industries as part of decision making
process either along a new product development, manufacturing process and improvement. It is not used only
in engineering areas it has been used in administration, marketing, hospitals, pharmaceutical [2], food industry
[3], energy and architecture [4][5], and chromatography [6]. DOE is applicable to physical processes as well
as computer simulation models [7].
carefully. In the 1920s and 1930s Ronald A. Fisher conducted a research in agriculture with the aim of
increasing yield of crop in the UK. Getting data and was challenging e.g. if he relayed on his traditional
method ANOVA (F-test, means Fisher - test) he may plant a crop in spring and get results in fall which is too
long for getting data. Finally, he came up with design of experiment and officially he was the first one who
started using DOE. In 1935, he wrote a book on DOE, in which he explained how valid conclusion could be
drawn from the experiment in presence of nuisance factors. He analyzed presence of nuisance factors with
fluctuation of weather conditions (temperature, rainfall, soil condition). Credit for Response Surface Method
(RSM) belongs to George Box who is also from the UK. He was concerned with experimental design
procedures for process optimization. In 1550s, W. Edwards Deming was concerned with design of
experiment as well as statistical methods. Genichi Taguchi was Japanese statistician concerned with quality
improvement methods. He contributed to statistic by introducing Loss function and experiments extending
with an "outer array" in DOE as an advanced method in the Six Sigma initiatives [8].
422
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
90000
80000
70000
60000
Publications
50000
40000
30000
20000
10000
Year
100000
90000 y = 2960.3x - 6E+06
80000 R² = 0.9708
70000
Publications
60000
50000
40000
30000
20000
10000
0
Year
Figure 2. Progressive use of DOE as scientific method over past two decades[13]2
1
Data obtained from Scopus for search “design of experiments" OR "experimental design" OR "DOE" in Title ‒ abstact ‒ Key words
2
Data obtained from Scopus for search “design of experiments" OR "experimental design" OR "DOE" in Title ‒ abstact ‒ Key words
423
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
DOE application in the future can be predicted with linear regression model, which is based on past date over
18 years. Therefore, it can be expectedthat DOE usage expansion will continue in the future including its
application in the existing and new scientific areas. The state of the art of DOE application in certain scientific
areas is shown in Figure 3.
Medicine 18%
Engineering 10%
Biochemistry & Genetics 10%
Physics & Astronomy 7%
Computer Science 6%
Social Science 5%
Materials 4%
Agricultural 4%
Chemistry 4%
Mathematics 4%
Earth & Planetary 3%
Environmental Science 3%
Arts & Humanities 3%
Neuroscience 2%
Immunology & Microbiology 2%
Chemical Engineering 2%
Pharmacology 2%
Psychology 2%
Business Management 2%
Economics Econometrics 2%
Energy 1%
Nursing 1%
Health Professions 1%
Decision Sciences 1%
Multidisciplinary 1%
Veterinary 0.4%
Dentistry 0.3%
Undefined 0.2%
3
Data obtained from Scopus for search "design of experiments" OR "experimental design" OR "DOE" in Title ‒ abstact ‒ Key words
424
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
4. Determine Experimental design type‒ e. g. a screening design is needed for significant factors
identification; or for optimization factor-response function is going to be planned, number of test
samples determination.
5. Perform experiment using design matrix.
6. Data analysis using statistical methods such as regression and ANOVA.
7. Practical conclusions and recommendations‒ including graphical representation of the results
and validation of the results.
Most difficult part of DOE is to plan experiment in term of selecting appropriate factors for testing
(x‒variables), what ranges of x's to select, how many replicates is supposed to be used, and is center point
required?
2. Statistical Background
H0: μ1 = μ1 = ... = μk = 0
(1)
H1: μ j ≠ 0 for at least one j different than zero.
The procedure of test involves an analysis of variance (ANOVA) and performing F-test. Observed value is
calculated as the ratio between treatment mean squares ( ) and error mean squares (error variance):
(2)
425
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
samples), n is number of observation for particular treatment. Total sum of squares ( is addition of sum of
squares of treatment and sum of squares of error and it is calculated as:
(3)
whre, is the predicted value for the ith test, is the mean of response variable observations, is the ith
observed value of the response variable.
Having calculated Fo, H0 can be accepted or rejected in the following cases:
(4)
H0 is going to be rejected if observed value of is grater than its critical value . The critical
value is taken from corresponding statistical table for significance level α, the degrees of freedom for the
numerator (a ‒ 1), and degrees of freedom for the denominator a(n ‒ 1). Also, H 0 is going to be accepted if
observed value is lower than its critical value ( ). If p-value approach for the statistics is
used than H0 is going to be rejected if p < α, and will be accepted if p > α.
(5)
where, y is response variable also called dependant variable, βj, βj, j = 0. 1, 2, …, k, are regression coefficients
or parameters and represent expected change in response variable per unit change in input variable, xj, is
regressesor variable or input variable or independent variable, ε is random error, which supposed to be
normally distributed with N (0, σ2) for given values of x and y.
Parameter estimate in multiple linear regression models is done using least squares method. In case that there
are multiple observations (n) on the response variable y1, y2, …yn, and that there are observation at each input
variable xij, (i = 1, 2, …, n) than it can be represented as matrix in Table 1 an written as[16]:
(6)
(7)
426
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
Referring to Table 1, it can be formulated using the general multivariate linear regression model in matrix
form as follows:
(8)
(9)
After slowing the function that minimizes sum of squared errors ε2, the least squared estimator of can be
calculated from the following:
(10)
Significance test of regression model is to determine existence of the significant relationship between input
variable and output variable. In this case it is tested hypothesis assuming that a least one parameter differs
from the others.
H0: β1 = β1 = ... βk = 0
(11)
H1: βj ≠ 0 for at least one j different than zero.
The procedure of test involves an analysis of variance (ANOVA) and performing F-test. Observed value is
calculated as the ratio between regression mean squares ( ) and error mean squares (error variance):
(12)
where, regression sum of squares, is error sum of squares,k is degree of freedom for the numerator, (n
‒ k ‒ 1) is degree of freedom for the denominator. Total sum of squares is addition of regression sum of
squares and error sum of squares and it is calculated as:
(13)
whre, is the predicted value for the ith test, is the mean of all response variable observations, is the ith
observed value of the response variable.
(14)
427
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
H0 is going to be rejected if observed value of is grater than its critical value . The critical value
is taken from corresponding statistical table for significance level α, the degrees of fridom for the numerator k,
and degrees of freedom for the denominator (n ‒ k ‒ 1). Also, H0 is going to be accepted if observed value is
lower than its critical value ( ). If p-value approach for the statistics is used than H0 is going
to be rejected if p < α, and will be accepted if p > α.
Goodness of fit of a model is measured by coefficient of determination R2, which is a measure of how well
real data points are approximated with regression model. It measures how amount of reductions of variability
in response variable caused using input variables in regression model.
(15)
Higher value of R2 means better fit. If R2 = 1 it means that there is perfect fit. However, having a higher value
dos not meat that there is a good fit and that regression model is good one because adding a new variable to
the model (either the variable is significant or not) will increase R2 value, which will lead to poor prediction.
To solve this, an adjusted R2 is introduced, which will not always increase with adding a new variable.
(16)
If non-significant variable is added to the model will decrease and vice versa.
3. Factorial Design
Factorial experiments can be design with one, two, three and more factors. Experiments with only one factor
are often called simple comparative experiments. In these cases, t‒test or ANOVA were used for analysis.
Factorial experiments with two factors (A and B) usually include two level factorial designs for identification
of factor effects on the response variable by investigating all possible combinations of the factor levels. The
factor effect is defined as change in the response variable by changing the level of the factor.
Factorial experiments with multiple factors (A, B, …, K), with two levels ("low" and "high") the complexity of
experimentation might be a problem. The number of possible combinations goes up with the number of
factors, for instance a 2-level design with 8 factor has 256 combination which very set such type of
experiments and analyze data. Multiple factor experiment requires a lot of resources, materials and it is time
consuming and expensive. Additional problem with multiple factorial design is to maintain experimental
conditions unchanged during a huge number of experiments.
Trying to overcome the problems with multiple factor factorial designs and depending from a case to case, it
is possible to be designed as Full Factorial Design 2k or Fractional Factorial Design 2k‒p. In this case number 2
represents number of levels, while k is number of factors and p is the fraction size of the full factorial used.
More details about full factorial and fractional factorial design with examples is provided below[17].
the number of replicates. For example, if three units / samples were tested at each treatment, the number of
replicates is three.
Design matrix for a 22 factorial design is given inTable 2:
(17)
Where, x1 and x2 represent coded variables of factors A and B respectively, while x1x2 cross‒product and
represents AB interaction, xj= ‒1 and xj= +1 represent low and high levels of factors.
(18)
Whre, contrast is obtained from Table 2 as sum of products of signs in corresponding column and the runs
listed in corresponding rows.
interactions, …, one k‒factor interaction. The procedure for a 2k factorial design was the following:
1. estimated factor effect ‒ effects are estimated and their magnitudes were examined with the aim of
important factors identification;
2. initial model formulation ‒ full model is included that takes in account all main effects and
interactions;
3. statistical testing ‒ ANOVA is used to test significance of main effects and interactions;
4. model refinement ‒ non-significant factors from initial model are removed;
5. residual analysis ‒ to check adequacy of the model and assumptions;
6. result interpretation ‒ graphical analysis of the results such as main effects, interactions etc.
General form of ANOVA mentioned in step 3 is shown in the following table [17]:
429
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
To estimate sum of squares and effects, contrasts associated with the effect were determined first.
(19)
(20)
(21)
430
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
(22)
where, is an observation for factor A at i-th level (i = 1, 2, … a) and for B factor at j-th level (j = 1, 2, …
b) and for k-th replicate (k = 1, 2, … n).
Sum of squares of error can be obtained from using the following equation:
(23)
I A B AB C AC BC ABC
(1) 1 -1 -1 1 -1 1 1 -1
a 1 1 -1 -1 -1 -1 1 1
b 1 -1 1 -1 -1 1 -1 1
ab 1 1 1 1 -1 -1 -1 -1
c 1 -1 -1 1 1 -1 -1 1
ac 1 1 -1 -1 1 1 -1 -1
bc 1 -1 1 -1 1 -1 1 -1
abc 1 1 1 1 1 1 1 1
431
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
Table 5. The 23‒1 factorial design factorial design matrix with the defining relations I = ABC and I = ‒ABC
I A B AB C AC BC ABC
a 1 1 -1 -1 -1 -1 1 1
b 1 -1 1 -1 -1 1 -1 1
c 1 -1 -1 1 1 -1 -1 1
abc 1 1 1 1 1 1 1 1
(1) 1 -1 -1 1 -1 1 1 -1
ab 1 1 1 1 -1 -1 -1 -1
ac 1 1 -1 -1 1 1 -1 -1
bc 1 -1 1 -1 1 -1 1 -1
abc bc
c
+ + ac
C C
b ab
+ +
B B
‒ ‒ ‒ ‒
A a (1) A
(a) (b)
3‒1
Figure 4. The 2 factorial design factorial design orthogonal representation with defining relations: (a)
principal fraction I = ABC; (c) alternative fraction I = ‒ABC
In this example ABC is called the generator. It noticed that some columns in
Table 5 have the same values (they are the same) and they are colored in the same color (I = ABC, A = BC, B
= AC, C = AB), where I is called defining relation, which may be aliased with several effects. The alias can be
determined by using the defining relation I = ABC by multiplying any column by the defining relation yields
the aliases such as: A · I = A · ABC = A2BC . Since square of any column is equal to identity I, then it can be
written as: A = BC. The other aliases may be obtained, B · I = B · ABC = AB2C and C · I = C · ABC = ABC2,
and replacing squared member by I it is obtained B = AC, C = AB. Therefore, when A, B and C is estimated it
is actuality estimated A + BC, B + AC, C + AB.
Resolution a two-level fractional factorial design 23‒1 is called a resolution III design andit is equal to the
number of letters in the shortest word in the defining relation in which Roman numbers in subscript are used
to define resolution (e.g. design for I = ABC and I = ‒ABC; design for I = ABCD and I = ‒ABCD
etc.).
Designs with a resolution less than III are never used. Fractional factorial designs with the highest available
resolution with their defining relations is shown in Table 6.
432
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
Table 6. The 23‒1 factorial design factorial design matrix with the defining relations I = ABC and I = ‒ABC
Number of factors Fraction Resolution Runs (treatments) Defining relation
3 23‒1 III 4 I = ABC
4 24‒1 IV 8 I = ABCD
5 25‒1 V 16 I = ABCDE
6 26‒1 VI 32 I = ABCDEF
7 27‒1 VII 64 I = ABCDEFG
4. Application
The design matrix for 24 factorial design and single replicate response data for flash size in millimeters is
shown in Table 8.
Table 8. Design matrix for flash formation
Run Run A B C D AB AC AD BC BD CD ABC BC AC AB ABCD Flash
label D D D (mm)
1 (1) -1 -1 -1 -1 1 1 1 1 1 1 -1 -1 -1 -1 1 0.22
2 a 1 -1 -1 -1 -1 -1 -1 1 1 1 1 -1 1 1 -1 6.18
3 b -1 1 -1 -1 -1 1 1 -1 -1 1 1 1 -1 1 -1 0
4 ab 1 1 -1 -1 1 -1 -1 -1 -1 1 -1 1 1 -1 1 5.91
5 c -1 -1 1 -1 1 -1 1 -1 1 -1 1 1 1 -1 -1 6.6
6 ac 1 -1 1 -1 -1 1 -1 -1 1 -1 -1 1 -1 1 1 6.05
7 bc -1 1 1 -1 -1 -1 1 1 -1 -1 -1 -1 1 1 1 6.76
8 abc 1 1 1 -1 1 1 -1 1 -1 -1 1 -1 -1 -1 -1 8.65
9 d -1 -1 -1 1 1 1 -1 1 -1 -1 -1 1 1 1 -1 0.46
10 ad 1 -1 -1 1 -1 -1 1 1 -1 -1 1 1 -1 -1 1 5.06
11 bd -1 1 -1 1 -1 1 -1 -1 1 -1 1 -1 1 -1 1 0.55
12 abd 1 1 -1 1 1 -1 1 -1 1 -1 -1 -1 -1 1 -1 4.84
13 cd -1 -1 1 1 1 -1 -1 -1 -1 1 1 -1 -1 1 1 11.55
14 acd 1 -1 1 1 -1 1 1 -1 -1 1 -1 -1 1 -1 -1 9.9
15 bcd -1 1 1 1 -1 -1 -1 1 1 1 -1 1 -1 -1 -1 9.9
16 abcd 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 9.9
433
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
Applying Equations (19) and Equation (20) in the design matrix (Table 8) main effects are calculated and
shown in Table 9.
Table 9. Factor effect estimate and sum of squares
Model Effects Coefficients = SS Percent contribution
term (effect / 2)
A 2.5575 1.27 26.16323 0.12126637
B 0.061875 0.03 0.015314 7.0981E-05
C 5.76125 2.88 132.768 0.6153788
D 1.47125 0.73 8.658306 0.04013119
AB 0.466125 0.23 0.86909 0.00402823
AC -2.6345 -1.31 27.76236 0.12867835
AD -0.748 -0.374 2.238016 0.01037319
BC 0.218625 0.10 0.191188 0.00088615
BD -0.507375 -0.25 1.029718 0.00477273
CD 1.82325 0.91 13.29696 0.06163133
ABC 0.556875 0.27 1.240439 0.00574943
BCD -0.598125 -0.29 1.431014 0.00663274
ACD 2.22045E-16 1.11E-16 1.97E-31 9.1409E-34
ABD -0.130625 -0.06 0.068252 0.00031635
ABCD -0.067375 -0.03 0.018158 8.416E-05
Mean4 5.78
In this example there is a problem to apply F-test. Since there is only one replicate of the response variable F-
test is not possible to be done because error degrees of freedom is zero (MSE = SSE/0; ‒ division by zero is
undefined). Potential solution can be a graphical solution such as Normal probability diagram of effects 5, or
dropping entire factors from the model. Therefore based on Table 9, a graphical solution is used to represent
main effects and identification of those significant, which is shown in Figure 5.
120%
100% C
Normal % probability
A
80% CD
D
60%
40%
20%
AC
0%
-4 -2 0 2 4 6 8
effrcts
4
Mean response, calculated as average flash value.
434
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
From normal probability diagram of effects, it is observed that A, C, D, are significant factors and AC, and CD
are significant interactions. Plots of the main effects of factor A, C, D as well as AC and CD interactions are
shown in Figure 6.
6 8 6
6
4 4
4
2 2
2
0 0 0
0 1 2 3 0 1 2 3 0 1 2 3
a) b) c)
10 D‒ D+
C‒ C+
8 8
6 6
4
4
2
2
0
0 1 2 3 0
A 0 2A 4
d) e)
Figure 6. Main effects and interactions
Refereeing to Figure 6 it is observed that all three main effects are positive indicating that the molding process
supposed to be performed on low level to minimize flash size. Alongside main effects there are significant
interactions. Since main effect are involved with significant interactions than they have no much meaning.
Screw RPM and injection speed have little meaning at high pack pressure but has better performances in
reducing flash size at low level. Therefore, the best setup for getting better product quality (reducing flash
size), is obtained when all factors A, C, D are at the low level.
Since factor B is not significant it is dropped from the model. All associated interactions with B factor are
dropped from as well. Therefore, the design becomes a 23 factorial design with two replicates so called
hidden replication (still has 16 responses). In this the mean case sum of squares of error (MS E) can be
calculated because degrees of freedom are different than zero (2k(n‒1) = 23(2‒1) = 8). ANOVA result for 23
factorial design with two replicates is shown in Table 10.
435
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
Where, total sum of squares (SST) is obtained using Equation (22), model sum of squares is obtained as SSM =
Σ SSA,B,C…K = SSA + SSC + SSD + SSAD + SSCD, while error sum of squares is obtained as SSE = SST ‒ SSM.
Regression model that represent predicted/fitted value for flash formation is given below, and includes only
significant factor and interactions.
(24)
Based on Equation (24), it is possible to perform diagnostic check if it is applied to the residuals.
Table 11. Factor effect estimate and sum of squares
Run label y
(1) 0.22 -0.427625 0.647625
Normal probability plot of residuals and residuals versus predicted value of flash size is shown in Figure 7.
436
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
100% 2.5
2
Normal % probability
80% 1.5
1
60%
Residuals
0.5
40% 0
-0.5 40 60 80 100
20% -1
-1.5
0%
-2
-4.00 -2.00 0.00 2.00 4.00
-2.5
Residuals Predicterd
Residuals versus predicted value plot shows randomly dispersed data around the horizontal axis (predicted
values), therefore the variance has no tendency to change along predicted value axis indicating that linear
regression model is appropriate for the data.
label
Referring to Table 12 factors A, B and C, represent basic design factors that are aliased with three factor
interactions BCD, ACD and ABD respectively, while factor D is aliased with ABC three factor interaction.
The two factor interactions AB, AC and AD are re aliased with the other two factor interactions BC, BD and
CD respectively. Therefore, the relationships are: A = BCD, B = ACD, C = ABD, D = ABC, AB = BC, AC =
BD and AD = CD. Using Equation (20) and multiplying corresponding column with measured values (flash)
and summing up, a contrast is obtained for the certain column. Using Eq. (21), contrast is divided by
and main effects are obtained for number of replicates n = 1, and number of factors k = 3 and shown in Table
13.
437
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
1.00
0.90
C
0.80
AB
0.70
D
0.60
0.50 A
0.40
0.30
AC 0.20
0.10
0.00
-4 -2 0 2 4 6 8
5. Conclusion
Thanks to software development, the use of DOE as scientific tool has increased rapidly in past 20 years in
manufacturing and non‒manufacturing industries over the world. It was most popular tool used by scientist in
medicine (with 18%), engineering and biochemistry (with 20%), physics and computer science (with 13%),
providing about 50% participation of these scientific areas compared to the all other scientific areas. The trend
of use of DOE is rapidly growing and it is expected to slow down in the near future for current scientific
areas, but also will expanded over new scientific areas and have rapid grow there.
Implementing valid and efficient factorial experiments provide quantitative data that can be used as support
for decision making during system, process and product design or improvement. Full factorial design and
fractional factorial design with examples of application of DOE in product quality improvement is presented
with sep‒by‒step procedures and result interpretation. This can be a very useful guide to experimenter how to
design and conduct experiments, and how to analyze and interpret data.
438
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp.421–439
Refferences
[1] B. Duraković and H. Bašić, "Continuous Quality Improvement in Textile Processing by Statistical Process
Control Tools: A Case Study of Medium-Sized Company," Periodicals of Engineering and Natural
Sciences, vol. 1, no. 1, pp. 36-46, 2013.
[2] F. Paulo and L. Santos, "Design of experiments for microencapsulation applications: A review," Materials
Science and Engineering: C, vol. 77, no. August, pp. 1327-1340, 2017.
[3] P. Yu, M. Y. Low and W. Zhou, "Design of experiments and regression modelling in food flavour and
sensory analysis: A review," Trends in Food Science & Technology, vol. 71, no. January, pp. 202-215,
2018.
[4] A. Schlueter and P. Geyer, "Linking BIM and Design of Experiments to balance architectural and
technical design factors for energy performance," Automation in Construction, vol. 86, no. February , pp.
33-43, 2018.
[5] B. Durakovic and M. Torlak, "Simulation and experimental validation of phase change material and
water used as heat storage medium in window applications," J. of Mater. and Environ. Sci., vol. 8, no. 5,
pp. 1837-1746, 2017.
[6] D. B. Hibbert, "Experimental design in chromatography: A tutorial review," Journal of Chromatography
B, vol. 910, no. Dec, pp. 2-13, 2012.
[7] S. S. Garud, I. A. Karimi and M. Kraft, "Design of computer experiments: A review," Computers &
Chemical Engineering, vol. 106, no. November, pp. 71-95, 2017.
[8] J. K. Telford, "A Brief Introduction to Design of Experiments," Johns Hopkins APL Technical Digest, vol.
27, no. 3, pp. 224-232, 2007.
[9] B. Durakovic, H. Bašić and H. Muhič, "The Interrelationships between quality managment practicies and
their effects on innovation," in Trends in the Development of Machinery and Associated Technology,
Budapest, 2014.
[10] H. Guo and A. Mettas, "Design of Experiments and Data Analysis," in 2012 Annual Reliability and
Maintainability Symposium.
[11] B. Duraković and H. Bašić, "Textile Cutting Process Optimization Model Based On Six Sigma
Methodology In A Medium-Sized Company," Journal of Trends in the Development of Machinery and
Associated Technology, vol. 16, no. 1, pp. 107-110, 2012.
[12] B. Durakovic and M. Torlak, "Experimental and numerical study of a PCM window model as a thermal
energy storage unit," International Journal of Low-Carbon Technologies, vol. 12, no. 3, p. 272–280, 2017.
[13] "Scopus," 2017. [Online]. Available: www.scopus.com. [Accessed September 2017].
[14] D. Granato, V. M. d. A. Calado and B. Jarvis, "Observations on the use of statistical methods in Food
Science and Technology," Food Research International, vol. 55, p. 137–149, 2014.
[15] D. C. Montgomery, Statistical Quality Control, John Wiley, 2009.
[16] D. C. Montgomery and G. C. Runger, Applied Statistics and Probability for Engineers, Willey, 2005.
[17] D. C. Montgomery, Design and Analysis of Experiments, 5th edition, John Wiley and Sons, Inc. New York,
2001., 8 ed., John Wiley & Sons, Inc, 2013.
[18] H. Bašić, B. Duraković and A. Softić, "Six Sigma Model Testing In Optimizing Medium-Sized Company
Production Process," Journal of Trends in the Development of Machinery and Associated Technology,
vol. 16, no. 1, pp. 103-106, 2012.
[19] V. Okatia, A. Behzadmehra and S. Farsad, "Analysis of a solar desalinator (humidification–
dehumidification cycle) including a compound system consisting of a solar humidifier and subsurface
condenser using DoE," Desalination, vol. 397, no. November, pp. 9-21, 2016.
439